US 20050086462 A1
At the inputs and/or outputs, memories are assigned to a reconfigurable module to achieve decoupling of internal data processing and in particular decoupling of the reconfiguration cycles from the external data streams (to/from peripherals, memories, etc.).
10. A method for decoupling data streams in a reconfigurable module, comprising:
decoupling module-internal data transfers from external data transfers via at least one FIFO memory.
11. A method for decoupling data streams in a reconfigurable module, comprising:
providing multiple independent address generators.
12. The method as recited in
configuring the address generators independently of data processing.
13. A method for programming a reconfigurable module, comprising:
breaking down, by a compiler, address computations into multiple configurations.
14. A device for data processing, comprising:
a field of reconfigurable modules; and
an interface module configured for an exchange of data to be processed between the field of reconfigurable modules and an external circuit configuration, wherein the interface module includes at least one FIFO memory in which data may be stored and from which data may be retrieved.
15. The device as recited in
a circuit configuration to determine data which will be needed in a future configuration and to preloading the determined data into the interface configuration.
16. The device according to
17. The device as recited in
18. The device as recited in
means for backing up data of a configuration which has previously been executed and is at least partially no longer active at the time of data retrieval, the configuration data being stored in an interface memory, the configuration data already canceled by reconfiguration of participating field elements.
19. The device as recited in
means for backing up data of a configuration which has previously been executed and is at least partially no longer active at the time of data retrieval, for at least one of preventing deletion, and overwriting of data already sent once but not yet known to have been satisfactorily received.
20. The device as recited in
at least one of multiple field elements and field elements grouped into configurations of addressable memories to store data at least one of non-locally and temporarily.
21. The device as recited in
The present invention relates to reconfigurable components in general, and in particular but not exclusively the decoupling of data processing within the reconfigurable component and/or within parts of the reconfigurable component and data streams, specifically both within the reconfigurable component and also to and from peripherals, mass memories, host processors, and the like (see, e.g., German Patent Application Nos. DE 101 10 530.4 and DE 102 02 044.2).
Memories are assigned to a reconfigurable module (VPU) at the inputs and/or outputs to achieve decoupling of internal data processing, the reconfiguration cycles in particular, from the external data streams (to/from peripherals, memories, etc.).
Reconfigurable architecture includes modules (VPUs) having a configurable function and/or interconnection, in particular integrated modules having a plurality of unidimensionally or multidimensionally positioned arithmetic and/or logic and/or analog and/or storage and/or internally/externally interconnecting modules, which are interconnected directly or via a bus system.
These generic modules include in particular systolic arrays, neural networks, multiprocessor systems, processors having a plurality of arithmetic units and/or logic cells and/or communication/peripheral cells (IO), interconnecting and networking modules such as crossbar switches, as well as conventional modules including FPGA, DPGA, Chameleon, XPUTER, etc. Reference is also made in particular in this context to the following patents and patent applications of the same applicant: P 44 16 881.0-53, DE 197 81 412.3, DE 197 81 483.2, DE 196 54 846.2-53, DE 196 54 593.5-53, DE 197 04 044.6-53, DE 198 80 129.7, DE 198 61 088.2-53, DE 199 80 312.9, PCT/DE 00/01869, DE 100 36 627.9-33, DE 100 28 397.7, DE 101 10 530.4, DE 101 11 014.6, PCT/EP 00/10516, EP 01 102 674.7, DE 196 51 075.9, DE 196 54 846.2, DE 196 54 593.5, DE 197 04 728.9, DE 198 07 872.2, DE 101 39 170.6, DE 199 26 538.0, DE 101 42 904.5, DE 101 10 530.4, DE 102 02 044.2, DE 102 06 857.7, DE 101 35 210.7, EP 02 001 331.4, EP 01 129 923.7 as well as the particular parallel patent applications thereto. The entire disclosure of these documents are incorporated herein by reference.
The above-mentioned architecture is used as an example to illustrate the present invention and is referred to hereinafter as VPU. The architecture includes an arbitrary number of arithmetic, logic (including memory) and/or memory cells and/or networking cells and/or communication/peripheral (IO) cells (PAEs—Processing Array Elements), which may be positioned to form a unidimensional or multidimensional matrix (PA); the matrix may have different cells of any desired configuration. Bus systems are also understood here as cells. A configuration unit (CT) which affects the interconnection and function of the PA is assigned to the entire matrix or parts thereof.
Memory access methods for reconfigurable modules which operate according to a DMA principle are described in German Patent No. P 44 16 881.0, where one or more DMAs are formed by configuration. In German Patent Application No. 196 54 595.1, DMAs are fixedly implemented in the interface modules and may be triggered by the PA or the CT.
German Patent Application No. DE 196 54 846.2 describes how internal memories are written by external data streams and data is read out of the memory back into external units.
German Patent Application No. DE 199 26 538.0 describes expanded memory concepts according to DE 196 54 846.2 for achieving more efficient and easier-to-program data transmission. U.S. Pat. No. 6,347,346 describes a memory system which corresponds in all essential points to German Patent Application No. DE 196 54 846.2, having an explicit bus (global system port) to a global memory. U.S. Pat. No. 6,341,318 describes a method for decoupling external data streams from internal data processing by using a double-buffer method, in which one buffer records/reads out the external data while another buffer records/reads out the internal data; as soon as the buffers are full/empty, depending on their function, the buffers are switched, i.e., the buffer formerly responsible for the internal data now sends its data to the periphery (or reads new data from the periphery) and the buffer formerly responsible for the external data now sends its data to the PA (reads new data from the PA). These double buffers are used in the application to buffer a cohesive data area.
Such double-buffer configurations have enormous disadvantages in the data-stream area in particular, i.e., in data streaming, in which large volumes of data streaming successively into a processor field or the like must always be processed in the same way.
An object of the present invention is to provide a novel approach for commercial use.
A method according to an example embodiment of the present invention, in contrast to the previously known related art, allows a significantly simpler means of controlling the buffers, i.e., memories, connected in between; the related art is disadvantageous in the core area of typical applications of reconfigurable processors in particular. External and internal bus systems may be operated at different transfer rates and/or clock frequencies with no problem due to the memory devices connected in between because data is stored temporarily by the buffers. In comparison with inferior designs from the related art, this method requires fewer memory devices, typically only half as many buffers, i.e., data transfer interface memory devices, thus greatly reducing the hardware costs. The estimated reduction in hardware costs amounts to 25% to 50%. It is also simpler to generate addresses and to program the configuration because the buffers are transparent for the programmer. Hardware is simpler to write and to debug.
A paging method which buffers various data areas in particular for different configurations may be integrated.
It should first be pointed out that various memory systems are known as interfaces to the IO. Reference is made to German Patent No. and German Patent Application Nos. P 44 16 881.0, DE 196 54 595.1, and DE 199 26 538.0. In addition, a method is described in German Patent Applciation No. DE 196 54 846.2 in which data is first loaded from the IO, (1) data is stored within a VPU after being computed, (2) the array (PA) is reconfigured, (3) data is read out from the internal memory and written back to another internal memory, (4) this is continued until the fully computed result is sent to the IO. Reconfiguration means, for example, that a function executed by a part of the field of reconfigurable units or the entire field and/or the data network and/or data and/or constants which are necessary in data processing is/are determined anew. Depending on the application and/or embodiment, VPUs are reconfigured only completely or also partially, for example. Different reconfiguration methods are implementable, e.g., complete reconfiguration by switching memory areas (see, e.g., German Patent Application Nos. DE 196 51 075.9, DE 196 54 846.2) and/or wave reconfiguration (see, e.g., German Patent Application-Nos. DE 198 07 872.2, DE 199 26 538.0, DE 100 28 397.7, DE 102 06 857.7) and/or simple configuring of addressable configuration memories (see, e.g., German Patent Application Nos. DE 196 51 075.9, DE 196 54 846.2, DE 196 54 593.5). The entire disclosure of each of the particular patent specifications is expressly incorporated herewith.
In one example embodiment, a VPU is entirely or partially configurable by wave reconfiguration or by directly setting addressable configuration memories.
Thus, one of the main operating principles of VPU modules is to copy data back and forth between multiple memories, with additional and optionally the same operations (e.g., long FIR filter) and/or other operations (e.g., FFT followed by Viterbi) being performed with the same data during each copying operation. Depending on the particular application, data is read out from one or more memories and written into one or more memories.
For storing data streams and/or states (triggers, see, e.g., German Patent Application Nos. DE 197 04 728.9, DE 199 26 538.0), internal/external memories (e.g., as FIFOs) are used and corresponding address generators are utilized. Any appropriate memory architecture may be fixedly implemented specifically in the algorithm and/or flexibly configured.
For performance reasons, the internal memories of the VPU are preferably used, but basically external memories may also be used.
Assuming this, the following comments shall now be made regarding the basic design:
Interface modules which communicate data between the bus systems of the PA and external units are assigned to an array (PA) (see, e.g., German Patent No. P 44 16 881.0, and German Patent Application No. DE 196 54 595.1). Interface modules connect address buses and data buses in such a way as to form a fixed allocation between addresses and data. Interface modules may preferably generate addresses or parts of addresses independently.
Interface modules are assigned to FIFOs which decouple internal data processing from external data transmission. A FIFO here is a data-streamable buffer, i.e., input/output data memory, which need not be switched for data processing, in particular during execution of one and the same configuration. If other data-streamable buffers are known in addition to FIFO memories, they will subsequently also be covered by the term where applicable. In particular, ring memories having one or more pointers, in particular at least one write memory and one read memory, should also be mentioned. Thus, for example, during multiple reconfiguration cycles for processing an application, the external data stream may be maintained as largely constant, regardless of internal processing cycles. FIFOs are able to store incoming/outgoing data and/or addresses. FIFOs may be integrated into an interface module or assigned to one or more of them. Depending on the design, FIFOs may also be integrated into the interface modules, and at the same time additional FIFOs may be implemented separately. It is also possible to use data-streamable buffers integrated into the module, e.g., by integration of FIFO groups into a chip which forms a reconfigurable processor array.
In one example embodiment, multiplexers for free allocation of interface modules and FIFOs may also be present between the FIFOs (including those that are separate) and the interface modules. In one configuration, the connection of FIFOs to external modules or internal parts of the processor field performed by a multiplexer may be specified based on the processor field, e.g., by the PAE sending and/or receiving data, but it may also be determined, if desired, by a unit at a higher level of the hierarchy, such as a host processor in the case of division of data processing into a highly parallel part of the task and a poorly parallelizable part of the task and/or the multiplexer circuit may be determined by external specifications, which may be appropriate if, for example, it is indicated with the data which type of data is involved and how it is to be processed.
With regard to the external connection, units for protocol conversion between the internal and external bus protocols (e.g., RAMBUS, AMBA, PCI, etc.) are also provided. A plurality of different protocol converters may also be used within one embodiment. The protocol converters may be designed separately or integrated into the FIFOs or interface modules.
In one possible embodiment, multiplexers for free assignment of interface modules/FIFOs and protocol converters may be provided between the (separate) protocol converters and the interface modules/FIFOs. Downstream from the protocol converters there may be another multiplexer stage, so that a plurality of AMBA bus interfaces may be connected to the same AMBA bus, for example. This multiplexer stage may also be formed, for example, by the property of an external bus of being able to address a plurality of units.
In one example embodiment, the circuit operates in master and slave operating modes. In the master mode, addresses and bus accesses are generated by the circuit and/or the assigned PA; in slave mode, external units access the circuit, i.e., the PA.
In other embodiments, additional buffer memories or data collectors may be provided within the circuit, depending on the application, for exchanging data between interface modules. These buffer memories preferably operate in a random access mode and/or an MMU paging mode and/or a stack mode and may have their own address generators. The buffer memories are preferably designed as multi-port memories to permit simultaneous access of a plurality of interface modules. It is possible to access the buffer memories from a higher-level data processing unit, in particular from processors such as DSPs, CPUs, microcontrollers, etc., assigned to the reconfigurable module (VPU).
Now the decoupling of external data streams in particular will be described. According to one aspect of the present invention, the external data streams are decoupled by FIFOs (input/output FIFO, combined as IO-FIFO) which are used between protocol converters and interface modules.
The data processing method functions as follows: Through one or more input FIFOs, incoming data is decoupled from data processing in the array (PA). Data processing may be performed in the following steps:
With the preferred design of the input/output FIFOs (IO-FIFOs) as multi-ported FIFOs, data processing may be performed by protocol converters simultaneously with writing into and/or reading out from the particular FIFOs. The method described above yields a time decoupling which permits “quasi-steady-state” processing of constant data streams in such a way that there is only a latency but no interruption in the data stream when the first data packets have passed through. In an expanded embodiment, the IO-FIFOs may be designed so that the number of IO-FIFOs and their depth may be selected according to the application. In other words, IO-FIFOs may be distributed or combined (e.g., via a transmission gate, multiplexer/demultiplexer, etc.) so that there are more IO-FIFOs or they are deeper. For example, 8 FIFOs of 1,024 words each may be implemented and configured so that 8 FIFOs of 1,024 words or 2 FIFOs of 4,096 words are configured or, for example, 1 FIFO may be configured with 4,096 words and 4 with 1,024 words.
Modifications of the data processing method described here are possible, depending on the design of the system and the requirements of the algorithms.
In an expanded embodiment, the FIFOs function in such a way that in the case of output FIFOs the addresses belonging to the data inputs are also stored in the FIFOs and/or input FIFOs are designed so that there is one FIFO for the reading addresses to be sent out/already sent out and one FIFO for the incoming data words assigned to the addresses.
Below is a discussion of how a FIFO-RAM bank coupling, which is possible according to the present invention, may be implemented in a particularly preferred variant of the present invention.
Depending on the application, it is possible to conduct the data transfer with the IO-FIFOs via one or more additional memory stages (RAM bank) which are assigned locally to the array or are preferably coupled laterally to the array and only then relay data to the data processing PAEs (e.g., ALU-PAEs described in, e.g., German Patent Appliation No. DE 196 51 075.9).
In a preferred embodiment, RAM-PAEs have a plurality of data interfaces and address interfaces, they are thus designed as multi-port arrays. Designability of a data interface and/or address interface as a global system port should also be mentioned as a possibility.
Additional memory stage(s) (RAM banks) may be implemented, for example, by memory modules corresponding to the RAM-PAEs, as described in, for example, German Patent Application No. DE 196 54 846.2 and/or German Patent Application No. DE 199 26 538.0 and/or International Patent Application No. PCT/EP 00/10516.
In other words, a RAM-PAE may constitute a passive memory which is limited (essentially) to the memory function (see, e.g., German Patent Application No. DE 196 54 846.2) or an active memory which automatically generates and controls functions such as address computation and/or bus accesses (see, e.g., German Patent Application No. DE 199 26 538.0). In particular, in one possible embodiment, active address generation functions and/or data transfer functions may also be implemented for a “global system port.” Depending on the design, active memories may actively manage one or more data interfaces and address interfaces (active interfaces). Active interfaces may be implemented, for example, by additional modules such as sequencers/state machines and/or ALUs and/or registers, etc., within a RAM-PAE and/or by suitable hardwiring of an active interface to other PAEs whose function and networking are configured in one or more RAM-PAEs in accordance with the functions to be implemented. Different RAM-PAEs may be assigned to different other PAEs.
RAM-PAEs preferably have one or more of the following functions, i.e., modes of operation: random access, FIFO, stack, cache, MMU paging. In a preferred embodiment, RAM-PAEs are connected via a bus to a higher-level configuration unit (CT) and may be configured by it in their function and/or interconnection and/or memory depth and/or mode of operation. In addition, there is preferably also the possibility of preloading and reading out the memory contents by the CT, for example, to set constants and/or lookup tables (cos/sin).
Due to the use of multi-ported memories for the RAM-PAEs, writing and/or reading out of data into/from the IO-FIFOs and data access by the array (PA) may take place simultaneously, so that the RAM-PAEs may in turn again have a buffer property, as described in German Patent Application No. DE 196 54 846.2, for example.
RAM-PAEs may be combined (as discussed in International Patent Application No. PCT/EP 00/10516, for example) in such a way that larger memory blocks are formed and/or the RAM-PAEs operate so that the function of a larger memory is obtained (e.g., one 1,024-word RAM-PAE from two 512-word RAM-PAEs).
In an example embodiment, the units may be combined so that the same address is sent to multiple memories. The address is subdivided so that one portion addresses the entries in the memories and another portion indicates the number of the memory selected (SEL). Each memory has a unique number and may be selected unambiguously by comparing it with SEL. In a preferred embodiment, the number for each memory is configurable.
In another and/or additional example embodiment, an address is relayed from one memory to the next. This address is subdivided so that one portion addresses the entries in the memories and another portion indicates the number (SEL) of the memory selected. This is modified each time data is relayed; for example, a 1 may be subtracted from this each time data is relayed. The memory in which this address part has a certain value (e.g., zero) is activated.
In an example embodiment, the units may be combined so that the same address is sent to a plurality of memories. The address is subdivided so that one part addresses the entries in the memories and another part indicates the number (SEL) of the memory selected. A bus runs between memories, namely from one memory to the next, which has a reference address such that the address has a certain value (e.g., zero) in the first memory and this value is modified each time data is relayed (e.g., incremented by 1). Therefore, each memory has a different unique reference address. The portion of the address having the number of the selected memory is compared with the reference address in each case. If they are identical, the particular memory is selected. Depending on the design, the reference bus may be constructed using the ordinary data bus system or a separated bus.
In an example embodiment, there may be an area check of the address part SEL to rule out faulty addressing.
It should now be pointed out that RAM-PAEs may be used as FIFOs. This may be preferred in particular when a comparatively large memory capacity is provided by RAM-PAEs. Thus, in particular when using multi-ported memories for the RAM-PAEs, this yields the design option of dispensing with explicit IO-FIFOs and/or configuring a corresponding number of RAM-PAEs as FIFOs in addition to the IO-FIFOs and sending data from the IO to the corresponding memory ports. This embodiment may be regarded as particularly cost efficient because no additional memories need be provided, but instead the memories of the VPU architecture, which are configurable in their function and/or interconnection (see, e.g., German Patent Application No. DE 196 54 846.2, DE 199 26 538.0 and International Patent Application No. PCT/EP 00/10516), are configured corresponding to the character of configurable processors.
It is also possible to provide a multiplexer/demultiplexer upstream and/or downstream from the FIFO. Incoming and/or outgoing data streams may be formed from one or more data records. For example, the following function uses two incoming data streams (a and b) and one outgoing data stream (x):
This requirement may be met by using two approaches, for example:
Approach a) is supported according to the present invention by making available a sufficient number of IO channels and IO-FIFOs. However, this simple approach is unsatisfactory because an algorithm-dependent and very expensive number of IO channels, which cannot be determined precisely, must be made available.
Therefore, approach b) or a suitable combination of a) and b) may be preferred, e.g., two IO channels, one input and one output, data streams being multiplexed on each channel if necessary. It should be pointed out that the interfaces should be capable of processing data streams, i.e., a sufficiently high clock frequency and/or sufficiently short latencies should be provided on the internal and/or external buses. This may be the reason why a combination of the two variants may be particularly preferred, because by providing a plurality of parallel IO channels, the required clocking of external and/or internal buses may be reduced accordingly.
For approach b) or approaches based at least partially on approach b), it may be necessary to provide multiplexers and/or demultiplexers and to separate the data streams of one data channel (e.g., a) and b) should be separated from the input channel) or to combine a plurality of result channels on one output channel.
One or more multiplexers/demultiplexers (MuxDemux stage) may be located at different positions, depending on the technical hardware implementation and/or the functions to be executed. For example,
The MuxDemux stage may in turn either be fixedly implemented in the hardware and/or formed by a suitable configuration of any PAEs designed accordingly.
The position of the multiplexers/demultiplexers of the MuxDemux stage is determined by the configuration by a CT and/or the array (PA) and/or the IO itself, which may also be dynamically influenced, e.g., on the basis of the degree of filling of the FIFO(s) and/or on the basis of pending data transfers (arbitration).
In an example embodiment, the multiplexer/demultiplexer structure is formed by a configurable bus system (e.g., according to or resembling the bus system between the RAM/ALU/etc.—PAEs), whereby the bus system may in particular also be physically the same which is also used either by resource sharing or by a time multiplex method which may be implemented through a suitable reconfiguration.
It may be particularly preferred if addresses are generated in a particular manner, as is evident from the following discussion. Addresses for internal or external memories may be computed by address generators. For example, groups of PAEs may be configured accordingly and/or explicit address generators, implemented separately and specially, if necessary (e.g., DMAs such as those described in German Patent No. DE 44 16 881) or within interface cells (such as those described in German Patent Application No. DE 196 54 595.1) may be used. In other words, either fixedly implemented address generators, which are integrated into a VPU or are implemented externally, may be used and/or the addresses may be calculated by a configuration of PAEs according to the requirements of an algorithm.
Simple address generators are preferably fixedly implemented in the interface modules and/or active memories (e.g., RAM-PAES). For generation of complex address sequences (e.g., nonlinear, multidimensional, etc.), PAEs may be configured accordingly and connected to the interface cells. Such methods having the corresponding configurations are described in International Patent Application No. PCT/EP 00/10516.
Configured address generators may belong to another configuration (ConfigID, see, e.g., German Patent Application Nos. DE 198 07 872.2, DE 199 26 538.0 and DE 100 28 397.7) other than data processing. This makes a decoupling of address generation from data processing possible, so that in a preferred method, for example, addresses may already be generated and the corresponding data already loaded before or during the time when the data processing configuration is being configured. It should be pointed out that such data preloading and/or address pregeneration is particularly preferred for increasing processor performance, in particular by reducing latency and/or the wait clock cycle. Accordingly, the result data and its addresses may still be processed during or after removal of the data processing/generating configuration. In particular, it is possible through the use of memories and/or buffers such as the FIFOs described here, for example, to further decouple data processing from memory access and/or IO access.
In a preferred procedure, it may be particularly effective to combine fixedly implemented address generators (HARD-AG) (see, e.g., German Patent Application No. DE 196 54 595.1) and configurable address generators in the PA (SOFT-AG) in such a way that HARD-AGs are used for implementation of simple addressing schemes, while complex addressing sequences are computed by the SOFT-AG and then sent to the HARD-AG. In other words, individual address generators may overload and reset one another.
Interface modules for reconfigurable components are described in German Patent Application No. DE 196 54 595.1. The interface modules disclosed therein and their operation could still be improved further to increase processor efficiency and/or performance. Therefore, within the scope of the present invention, a particular embodiment of interface modules is proposed below such as that disclosed in particular in German Patent Application No. DE 196 54 595.1.
Each interface module may have its own unique identifier (IOID) which is transmitted from/to a protocol converter and is used for assigning data transfers to a certain interface module or for addressing a certain interface module. The IOID is preferably CT-configurable.
For example, the IOID may be used to select a certain interface module for a data transfer in the case of accesses by an external master. In addition, the IOID may be used to assign the correct interface module to incoming read data. To do so, the IOID is, for example, transmitted with the address of a data-read access to the IO-FIFOs and either stored there and/or relayed further to the external bus. IO-FIFOs assign the IOIDs of the addresses sent out to the incoming read data and/or the IOIDs are also transmitted via the external bus and assigned by external devices or memories to the read data sent back.
IOIDs may then address the multiplexers (e.g., upstream from the interface modules) so that they direct the incoming read data to the correct interface module.
Interface modules and/or protocol converters conventionally operate as bus masters. In a special embodiment, it is now proposed that interface modules and/or protocol converters shall function alternatively and/or fixedly and/or temporarily as bus slaves, in particular in a selectable manner, e.g., in response to certain events, states of state machines in PAEs, requirements of a central configuration administration unit (CT), etc. In an additional embodiment, the interface modules are expanded so that generated addresses, in particular addresses generated in SOFT-AGs, are assigned a certain data packet.
A preferred embodiment of an interface module is described below:
A preferred coupling of an interface module is accomplished by connecting any PAEs (RAM, ALU, etc.) and/or the array (PA) via a bus (preferably configurable) to interface modules which are either connected to the protocol converters or have the protocol converters integrated into them.
In a variant embodiment, IO-FIFOs are integrated into the interface modules.
For write access (the VPU sends data to external IOs, e.g., memories/peripherals, etc.) it is advantageous to link the address output to the data output, i.e., a data transfer takes place with the IO precisely when a valid address word and a valid data word are applied at the interface module, the two words may be originating from different sources. Validity may be identified by a handshake protocol (RDY/ACK) according to German Patent Application Nos. DE 196 51 075.9 or DE 101 10 530.4, for example. Through suitable logic gating (e.g., AND) of RDY signals of address word and data word, the presence of two valid words is detectable, and IO access may be executed. On execution of the IO access, the data words and the address words may be acknowledged by generating a corresponding ACK for the two transfers. The IO access including the address and data, as well as the associated status signals, if necessary, may be decoupled in output FIFOs according to the present invention. Bus control signals are preferably generated in the protocol converters.
For read access (the VPU receives data from external IOs, e.g., memories/peripherals, etc.), the addresses for the access are first generated by an address generator (HARD-AG and/or SOFT-AG) and the address transfer is executed. Read data may arrive in the same clock cycle or, at high frequencies, may arrive pipelined one or more clock cycles later. Both addresses and data may be decoupled through IO-FIFOs.
The conventional RDY/ACK protocol may be used for acknowledgment of the data, and it may also be pipelined (see, e.g., German Patent Application Nos. DE 196 54 595.1, DE 197 04 742.4, DE 199 26 538.0, DE 100 28 397.7 and DE 101 10 530.4).
The conventional RDY/ACK protocol may also be used for acknowledgment of the addresses. However, acknowledgment of the addresses by the receiver results in a very long latency, which may have a negative effect on the performance of VPUs. The latency may be bypassed in that the interface module acknowledges receipt of the address and synchronizes the incoming data assigned to the address with the address.
Acknowledgment and synchronization may be performed by any suitable acknowledgment circuit. Two possible embodiments are explained in greater detail below, although in a non-limiting fashion:
A FIFO stores the outgoing address cycles of the external bus transfers. With each incoming data word as a response to an external-bus access, the FIFO is instructed accordingly. Due to the FIFO character, the sequence of outgoing addresses corresponds to the sequence of outgoing data words. The depth of the FIFO (i.e., the number of possible entries) is preferably adapted to the latency of the external system, so that any outgoing address may be acknowledged without latency and optimum data throughput is achieved. Incoming data words are acknowledged according to the FIFO entry of the assigned address. If the FIFO is full, the external system is no longer able to accept any additional addresses and the current outgoing address is not acknowledged and is thus held until data words of a preceding bus transfer have been received and one FIFO entry has been removed. If the FIFO is empty, no valid bus transfer is executed and possibly incoming data words are not acknowledged.
b) Credit Counter
Each outgoing address of external bus transfers is acknowledged and added to a counter (credit counter). Incoming data words as a response to an external bus transfer are subtracted from the counter. If the counter reaches a defined maximum value, the external system can no longer accept any more addresses and the current outgoing address is not acknowledged and is thus held until data words of a preceding bus transfer have been received and the counter has been decremented. If the counter content is zero, no valid bus transfer is executed and incoming data words are not acknowledged.
To optimally support burst transfers, the method using a) (FIFO) is particularly preferred, and in particular FIFOs may be used like the FIFOs described below for handling burst accesses and the assignment of IOIDs to the read data.
The IO-FIFOs described here may be integrated into the interface modules. In particular, an IO-FIFO may also be used for embodiment variant a).
The optional possibility of providing protocol converters is discussed above. With regard to particularly advantageous possible embodiments of protocol converters, the following comments should be made:
A protocol converter is responsible for managing and controlling an external bus. The detailed structure and functioning of a protocol converter depend on the design of the external bus. For example, an AMBA bus requires a protocol converter different from a RAMBUS. Different protocol converters are connectable to the interface modules, and within one embodiment of a VPU, a plurality of, in particular, different protocol converters may be implemented.
In one preferred embodiment, the protocol converters are integrated into the IO-FIFOs of the present invention.
It is possible according to the present invention to provide burst bus access. Modern bus systems and SoC bus systems transmit large volumes of data via burst sequences. An address is first transmitted and data is then transmitted exclusively for a number of cycles (see AMBA Specification 2.0, ARM Limited).
For correctly executing burst accesses, several tasks are to be carried out:
1) Recognizing Burst Cycles
Linear bus accesses, which may be converted into bursts, must be recognized to trigger burst transfers on the external bus. For recognizing linear address sequences, a counter (TCOUNTER) may be used; it is first loaded with a first address of a first access and counts linearly up/down after each access. If the subsequent address corresponds to the counter content, there is a linear and burst-capable sequence.
2) Aborting at Boundaries
Some bus systems (e.g., AMBA) allow bursts (a) only up to a certain length and/or (b) only up to certain address limits (e.g., 1024 address blocks). For (a), a simple counter may be implemented according to the present invention, which counts from the first desired or necessary bus access the number of data transmissions and at a certain-value which corresponds to the maximum length of the burst transfer, signals the boundary limits using a comparator, for example. For L(b), the corresponding bit (e.g., the 10th bit for 1024 address limits) which represents the boundary limit may be compared between TCOUNTER and the current address (e.g., by an XOR function). If the bit in the TCOUNTER is not equal to the bit in the current address, there has been a transfer beyond a boundary limit which is signaled accordingly.
3) Defining the Length
If the external bus system does not require any information regarding the length of a burst cycle, it is possible and preferable according to the present invention to perform burst transfers of an indefinite length (cf. AMBA). If length information is expected and/or certain burst lengths are predetermined, the following procedure may be used according to the present invention. Data and addresses to be transmitted are written into a FIFO, preferably with the joint use of the IO-FIFO, and are known on the basis of the number of addresses in the (IO-)FIFO. For the addresses, an address FIFO is used, transmitting in master mode the addresses from the interface modules to the external bus and/or operating conversely in slave mode. Data is written into a data FIFO, which transmits data according to the transmission (read/write). In particular, a different FIFO may be used for write transfers and for read transfers. The bus transfers may then be subdivided into fixed burst lengths, so that they are known before the individual burst transfers and may be stated on initiation of the burst, burst transfers of the maximum burst length preferably being formed first and if the number of remaining (IO-)FIFO entries is smaller than the current burst length, a next smaller burst length is used in each case. For example, ten (IO-;)FIFO entries may be transmitted at a maximum burst length of 4 with 4, 4, 2 burst transfers.
4) Error Recovery
Many external bus systems (cf. AMBA) provide methods for error elimination in which failed bus transfers are repeated, for example. The information as to whether a bus transfer has failed is transmitted at the end of a bus transfer, more or less as an acknowledgment for the bus transfer. To repeat a bus transfer, it is now necessary for all the addresses to be available, and in the case of write access, the data to be written away must also be available. According to the present invention, the address FIFOs (preferably the address FIFOs of the IO-FIFOs) are modified so that the read pointer is stored before each burst transfer. Thus, a FIFO read pointer position memory means is provided, in particular an address FIFO read pointer position memory means. This may form an integral part of the address FIFO in which, for example, a flag is provided, indicating that information stored in the FIFO represents a read pointer position or it may be provided separately from the FIFO. As an alternative, a status indicating deletability could also be assigned to data stored in the FIFO, this status also being stored and reset to “deletable” if successful data transmission has been acknowledged. If an error has occurred, the read pointer is reset at the position stored previously and the burst transfer is repeated. If no error has occurred, the next burst transfer is executed and the read pointer is restored accordingly. To prevent the write pointer from arriving at a current burst transfer and thus overwriting values which might still be needed in a repeat of the burst transfer, the full status of the FIFOs is determined by comparing the stored read pointer with the write pointer.
IO-FIFOs and/or FIFOs for managing burst transfers may preferably be expanded to incoming read data using the function of address assignment, which is known from the interface modules. Incoming read data may also be assigned the IOID which is preferably stored in the FIFOs together with the addresses. Through the assignment of the IOID to incoming read data, the assignment of the read data to the corresponding interface modules is possible by switching the multiplexers according to the IOIDs, for example.
According to the present invention, it is possible to use certain bus systems and/or to design bus systems in different ways. This is described in further detail below. Depending on the design, different bus systems may be used between the individual units, in particular the interface modules, the IO-FIFOs, the protocol converters, and a different bus system may be implemented between each of two units. Different designs are implementable, the functions of a plurality of designs being combinable within one design. A few design options are described below.
The simplest possible design is a direct connection of two units.
In an expanded embodiment, multiplexers are provided between the units, which may have different designs. This example embodiment is preferred in particular when using a plurality of the particular units.
A multiplex function may be obtained using a configurable bus, which is configurable by a higher-level configuration unit (CT), specifically for a period of time for the connection of certain units.
In an example embodiment, the connections are defined by selectors which decode a portion of an address and/or an IOID, for example, by triggering the multiplexers for the interconnection of the units. In a particularly preferred embodiment, the selectors are designed in such a way that a plurality of units may select a different unit at the same time, each of the units being arbitrated for selection in chronological sequence. An example of a suitable bus system is described in, e.g., German Patent Application No. DE 199 26 538.0.
Additional states may be used for arbitration. For example, data transfers between the interface modules and the IO-FIFOs may be optimized as follows:
In each case one block of a defined size of data to be transmitted is combined within the FIFO stages. As soon as a block is full/empty, a bus access is signaled to the arbiter for transmitting the data. Data is transmitted in a type of burst transfer, i.e., the entire data block is transmitted by the arbiter during a bus allocation phase. In other words, a bus allocation may take place in a manner determined by FIFO states of the connected FIFOs, data blocks being used for the determination of state within a FIFO. If a FIFO is full, it may arbitrate the bus for emptying; if a FIFO is empty, it may arbitrate the bus for filling. Additional states may be provided, e.g., in flush, which is used for emptying only partially full FIFOs and/or for filling only partially empty FIFOs. For example, flush may be used in a change of configuration (reconfiguration).
In a preferred embodiment, the bus systems are designed as pipelines in order to achieve high data transfer rates and clock rates by using suitable register stages and may also function as FIFOs themselves, for example.
In a preferred embodiment, the multiplexer stage may also be designed as a pipeline.
According to the present invention, it is possible to connect a plurality of modules to one IO and to provide communication among the modules. In this regard, the following should be pointed out:
A plurality of these configuration modules may be configured simultaneously into the PA, dependently and/or independently of one another.
The configuration modules must be hardwired to a limited IO, which is typically provided in particular only at certain locations and is therefore not relocatable, in such a way that the configuration modules are able to use the IOs simultaneously and data is assigned to the correct modules.
In addition, configuration modules that belong together (dependent) must be hardwired together in such a way that free relocation of the configuration modules is possible among one another in the PA.
Such a flexible design is in most cases not possible through the conventional networks (see, e.g., German Patent Nos. P 44 16 881.0, 02, 03, 08), because this network must usually be explicitly allocated and routed through a router.
German Patent Application No. DE 197 04 742.4 describes a method of constructing flexible data channels within a PAE matrix according to the algorithms to be executed so that a direct connection through and in accordance with a data transmission is created and subsequently dismantled again. Data to be transmitted may be precisely assigned to one source and/or one destination.
In addition and/or as an alternative to German Patent Application No. DE 197 04 742.4 and the procedures and configurations described therein, additional possibilities are now provided through the present invention, and methods (hereinafter referred to jointly as GlobalTrack) that permit flexible allocation and interconnection during run time may be used, e.g., serial buses, parallel buses and fiber optics, each with suitable protocols (e.g., Ethernet, Firewire, USB). Reference is made here explicitly to transmission by light using a light-conducting substrate, in particular with appropriate modulation for decoupling of the channels. Another particular feature of the present invention with respect to memory addressing, in particular paging and MMU options, is described below.
Data channels of one or multiple GlobalTracks may be connected via mediating nodes to an ordinary network, e.g., according to German Patent Nos. P 44 16 881.0, 02, 03, 08. Depending on the implementation, the mediating nodes may be configured differently in the PA, e.g., assigned to each PAE, to a group and/or hierarchy of PAEs, and/or to every nth PAE.
In a particularly preferred embodiment, all PAEs, interface modules, etc., have a dedicated connection to a GlobalTrack.
A configuration module is designed in such a way that it has access to one or a plurality of these mediating nodes.
A plurality of configuration modules among one another and/or configuration modules and IOs may now be connected via the GlobalTrack. With proper implementation (e.g., German Patent Application No. DE 197 04 742.4) a plurality of connections may now be established and used simultaneously. The connection between transmitters and receivers may be established in an addressed manner to permit individual data transfer. In other words, transmitters and receivers are identifiable via GlobalTrack. An unambiguous assignment of transmitted data is thus possible.
Using an expanded IO, which also transmits the transmitter address and receiver address—as is described in German Patent Application No. DE 101 10 530.4, for example—and the multiplexing methods described in German Patent Application No. DE 196 54 595.1, data for different modules may be transmitted via the IO and may also be assigned unambiguously.
In a preferred embodiment, data transfer is synchronized by handshake signals, for example. In addition, data transfer may also be pipelined, i.e., via a plurality of registers implemented in the GlobalTrack or assigned to it. In a very complex design for large-scale VPUs or for their interconnection, a GlobalTrack may be designed in a network topology using switches and routers; for example, Ethernet could be used.
It should be pointed out that different media may be used for GlobalTrack topologies, e.g., the method described in German Patent Application No. DE 197 04 742.4 for VPU-internal connections and Ethernet for connections among VPUs.
Memories (e.g., RAM-PAEs) may be equipped with an MMU-like paging method. For example, a large external memory could then be broken down into segments (pages), which in the case of data access within a segment would be loaded into one of the internal memories and, at a later point in time, after termination of data access, would be written back into the external memory.
In a preferred embodiment, addresses sent to a (internal) memory are broken down into an address area, which is within the internal memory (MEMADR) (e.g., the lower 10 bits in a 1,024-entry memory) and a page address (the bits above the lower 10). The size of a page is thus determined by MEMADR.
The page address is compared with a register (page register) assigned to the internal memory. The register stores the value of the page address last transferred from a higher-level external (main) memory into the internal memory.
If the page address matches the page register, free access to the internal memory may take place. If the address does not match (page fault), the current page content is written, preferably linearly, into the external (main) memory at the location indicated by the page register.
The memory area in the external (main) memory (page) which begins at the location of the current new page address is written into the internal memory.
In a particularly preferred embodiment, it is possible to specify by configuration whether or not, in the event of a page fault, the new page is to be transferred from the external (main) memory into the internal memory.
In a particularly preferred embodiment, it is possible to specify by configuration whether or not, in the event of a page fault, the old page is to be transferred from the internal memory into the external (main) memory.
The comparison of the page address with the page register preferably takes place within the particular memory. Data transfer control in the event of page faults may be configured accordingly by any PAEs and/or may take place via DMAs (e.g., in the interface modules or external DMAs). In a particularly preferred embodiment, the internal memories are designed as active memories having integrated data transfer control (see, e.g., German Patent Application No. DE 199 26 538.0).
In another possible embodiment, an internal memory may have a plurality (p) of pages, the size of a page then preferably being equal to the size of the memory divided by p. A translation table (translation look-aside buffer=TLB) which is preferably designed like a fully associative cache replaces the page register and translates page addresses to addresses in the internal memory; in other words, a virtual address may be translated into a physical address. If a page is not included in the translation table (TLB), a page fault occurs. If the translation table has no room for new additional pages, pages may be transferred from the internal memory into the external (main) memory and removed from the translation table so that free space is again available in the internal memory.
It should be pointed out explicitly that a detailed discussion is not necessary because a plurality of conventional MMU methods may be used and may be used with only minor and obvious modifications.
The possibility of providing a collector memory, as it is known, has been mentioned above. In this regard, the following details should also be mentioned.
A collector memory (collector) capable of storing larger volumes of data may be connected between the interface modules and IO-FIFOs.
The collector may be used for exchanging data between the interface modules, i.e., between memories assigned to the array (e.g., RAM-PAEs).
The collector may be used as a buffer between data within a reconfigurable module and external data.
A collector may function as a buffer for data between different reconfiguration steps; for example, it may store data of different configurations while different configurations are active and are being configured. At deactivation of configurations, the collector stores their data, and data of the newly configured and active configurations is transmitted to the PA, e.g., to memories assigned to the array (RAM-PAEs).
A plurality of interface modules may have access to the collector and may manage data in separate and/or jointly accessible memory areas.
In a preferred embodiment, the collector may have multiple terminals for interface modules, which may be accessed simultaneously (i.e., it is designed as a multi-port collector device).
The collector has one or more terminals to an external memory and/or external peripherals. These terminals may be connected to the IO-FIFOs in particular.
In an expanded embodiment, processors assigned to the VPU, such as DSPs, CPUs and microcontrollers, may access the collector. This is preferably accomplished via another multi-port interface.
In a preferred embodiment, an address translation table is assigned to the collector. Each interface may have its own address translation table or all the interfaces may share one address translation table. The address translation table may be managed by the PA and/or a CT and/or an external unit. The address translation table is used to assign collector memory areas to any addresses and it operates like an MMU system. If an address area (page) is not present within the collector (pagemiss), this address area may be loaded into the collector from an external memory. In addition, address areas (pages) may be written from the collector into the external memory.
For data transfer to or between the external memory, a DMA is preferably used. A memory area within the collector may be indicated to the DMA for reading or writing transmission; the corresponding addresses in the external memory may be indicated separately or preferably removed by the DMA from the address translation table.
A collector and its address generators (e.g., DMAs) may preferably operate according to or like MMU systems, which are conventional for processors according to the related art. Addresses may be translated by using translation tables (TLB) for access to the collector. According to the present invention, all MMU embodiments and methods described for internal memories may also be used on a collector. The operational specifics will not be discussed further here because they correspond to or closely resemble the related art.
In an expanded or preferred embodiment, a plurality of collectors may be implemented.
According to the present invention, it is possible to optimize access to memory. The following should be pointed out in this regard:
One basic property of the preferred reconfigurable VPU architecture PACT-XPP is the possibility of superimposing reconfiguration and data processing (see, e.g., German Patent No. P 44 16 881.0, and German Patent Application Nos. DE 196 51 075.9, DE 196 54 846.2, DE 196 54 593.5, DE 198 07 872.2, DE 199 26 538.0, DE 100 28 397.7, DE 102 06 857.7). In other words, for example:
Modern memory protocols (e.g., SDRAM, DDRAM, RAMBUS) usually have the following sequence or a sequence having a similar effect, but steps 2 and 3 may possibly also occur in the opposite order:
This property may be utilized in a performance-efficient manner in VPU technology. For example, it is possible to separate the steps of computation of the address(es), initialization of memory access, data transfer and data processing in the array (PA) in such a way that different (chronological) configurations occur, so that largely optimum superpositioning of the memory cycles and data processing cycles may be achieved. Multiple steps may also be combined, depending on the application.
For example, the following method corresponds to this principle:
The application AP, which includes a plurality of configurations (ap=1, 2, . . . , z), is to be executed. Furthermore, additional applications/configurations which are combined under WA are to be executed on the VPU:
Between the steps, any configuration from WA may be executed, e.g., when a waiting time is necessary between steps, because data is not yet available.
Likewise, in parallel with the processing of AP, configurations from WA may be executed during the steps, e.g., if AP does not use the resources required for WA.
It will be self-evident to those skilled in the art that variously modified embodiments of this method are also possible.
In one possible embodiment, the processing method may take place as shown below (Z marks a configuration cycle, i.e., a unit of time):
This sequence may be utilized efficiently by the data processing method described in, for example, German Patent Application No. DE 102 02 044.2 in particular.
The methods and devices described above are preferably operated using special compilers, which are expanded in particular in comparison with traditional compilers. The following should be pointed out in this regard:
For generating configurations, compilers that run on any computer system are used. Typical compilers include, for example, C-compilers and/or even NML compilers for VPU technology, for example. Particularly suitable compiler methods are described in German Patent Application Nos. DE 101 39 170.6, and DE 101 29 237.6, and European Patent No. EP 02 001 331.4, for example.
The compiler, at least partially, preferably takes into account the following particular factors: Separation of addressing into
Bus transfers are broken down into internal and external transfers.
bt1, bt2, and bt3 may be translated into different configurations which may, if necessary, be executed at a different point in time.
This method will now be illustrated on the basis of the following example:
function example (a, b:integer)→x:integer
This function is transformed by the compiler into three parts, i.e., configurations (subconfig):
example#dstore: Writes the results from the internal memory into externally (memories, peripherals, etc.).
An effect of the example method is that instead of i*j=100*100=10,000 external accesses, only i+j=100+100=200 external accesses are performed for reading the operands. These accesses are also completely linear, which greatly accelerates the transfer rate in modern bus systems (burst) and/or memories (SDRAM, DDRAM, RAMBUS, etc.).
Internal memory accesses take place in parallel, because different memories have been assigned to the operands.
For writing the results, i=100 external accesses are necessary and may again be performed linearly at maximum performance.
If the number of data transfers is not known in advance (e.g., WHILE loop) or is very large, a method may be used which reloads the operands as necessary through subprogram call instructions and/or writes the results externally. In a preferred embodiment, the states of the FIFOs may (also) be queried: “empty” if the FIFO is empty and “full” if the FIFO is full. The program flow responds according to the states. It should be pointed out that certain variables (e.g., ai, bi, xi) are defined globally. For performance optimization, a scheduler may execute the configurations example#dloada, example#dloadb before calling up example#process according to the methods already described, so that data is already preloaded. Likewise, example#dstore(n) may still be called up after termination of example#process in order to empty r#x.
The subprogram call instructions and managing of the global variables are comparatively complex for reconfigurable architectures. Therefore, in a preferred embodiment, the following optimization may be performed; in this optimized method, all configurations are run largely independently and are terminated after being completely processed (terminate). Since data b[j] is required repeatedly, example#dloadb must accordingly be run through repeatedly. To do so, for example, two alternatives will be described:
While “idle,” a configuration is inactive (waiting).
To avoid waiting cycles, configurations may also be terminated as soon as they are temporarily no longer able to continue fulfilling their function. The corresponding configuration is removed from the reconfigurable module but remains in the scheduler. Therefore, the “reenter” instruction is used for this below. The relevant variables are saved before termination and are restored when configuration is repeated:
With regard to the preceding discussion and to the following, the possibility of using a ‘context switch’ according to the present invention should also be pointed out. In this regard, the following should be noted:
Repeated start of configurations, e.g., “reenter,” requires that local data (e.g., ai, bi, xi) be backed up and restored. Known related-art methods provide explicit interfaces to memories or to a CT to transmit data. All of these methods may be inconsistent and/or may require additional hardware.
The context switch according to the present invention is implemented in such a way that a first configuration is removed; data to be backed up remains in the corresponding memories (REGs) (memories, registers, counters, etc.).
A second configuration is loaded; this connects the REGs in a suitable manner and in a defined sequence to one or multiple global memory (memories).
The configuration may use address generators, for example, to access the global memory (memories).
The configuration may use address generators, for example, to access REGs designed as memories.
According to the configured connection between the REGs, the contents of the REGs are written into the global memory in a defined sequence, the particular addresses being predetermined by address generators. The address generator generates the addresses for the global memory (memories) in such a way that the memory areas (PUSHAREA) that have been written are unambiguously assigned to the first configuration removed.
In other words, different address areas are preferably provided for different configurations.
The configuration corresponds to a PUSH of ordinary processors.
Other configurations subsequently use the resources.
The first configuration is to be started again, but first a third configuration which connects the REGs of the first configuration in a defined sequence is started.
The configuration may use address generators, for example, to access the global memory or memories. The configuration may use address generators, for example, to access REGs designed as memories.
An address generator generates addresses, so that correct access to the PUSHAREA assigned to the first configuration takes place. The generated addresses and the configured sequence of the REGs are such that data of the REGs is written from the memories into the REGs in the original order. The configuration corresponds to a POP of ordinary processors.
The first configuration is restarted.
In summary, a context switch is implemented in such a way that data to be backed up is exchanged with a global memory by loading particular configurations which operate like processor architectures known from PUSH/POP.
There is also the possibility of providing a special task switch and/or multiconfiguration handling.
In a preferred mode of operation, different data blocks of different configurations may be partitioned. These partitions may be accessed in a time-optimized manner by preloading a portion of the operands of a subsequent configuration P from external (main) memories and/or other (peripheral) data streams into the internal memories, e.g., during execution of a configuration Q, and during the execution of P, the results of Q as a portion of the total result from the internal memories are written into external (main) memories and/or other (peripheral) data streams.
The functioning here differs considerably from that described in, for example, U.S. Pat. No. 6,341,318. A data stream or data block is preferably decoupled by a FIFO structure (e.g., IO-FIFO). Different data streams or data blocks of different configurations in particular are preferably decoupled by different memories and/or FIFO areas and/or assignment marks in the FIFOs.
The optional MMU methods described above may be used for decoupling and buffering external data. In one type of application, a large external data block may be broken down into a plurality of segments, each may be processed within a VPU.
In an additional preferred mode of operation, different data blocks of different configurations may be broken down into partitions according to the method described above, these partitions now being defined as pages for an MMU. In this way, time-optimized access is possible by preloading the operands of a subsequent configuration P as a page from external (main) memories and/or other (peripheral) data streams into the internal memories, e.g., during execution of a configuration Q in the PA, and during the execution of P, the results of Q as a page from the internal memories are written into external (main) memories and/or other (peripheral) data streams.
For the methods described above, preferably internal memories capable of managing a plurality of partitions and/or pages are used.
These methods may be used for RAM-PAEs and/or collector memories.
Memories having a plurality of bus interfaces (multi-port) are preferably used to permit simultaneous access of MMUs and/or the PA and/or additional address generators/data transfer devices.
In one embodiment, identifiers are also transmitted in the data transfers, permitting an assignment of data to a resource and/or an application. For example, the method described in German Patent Application No. DE 101 10 530.4 may be used. Different identifiers may also be used simnultaneously.
In a particularly preferred embodiment, an application identifier (APID) is also transmitted in each data transfer along with the addresses and/or data. An application includes a plurality of configurations. On the basis of the APID, the transmitted data is assigned to an application and/or to the memories or other resources (e.g., PAEs, buses, etc.) intended for an application. To this end, the APIDs may be used in different ways.
Interface modules, for example, may be selected by APIDs accordingly.
Memories, for example, may be selected by APIDs accordingly.
PAEs, for example, may be selected by APIDs accordingly.
For example, memory segments in internal memories (e.g., RAM-PAEs, collector(s)) may be assigned by APIDs. To do so, the APIDS, like an address part, may be entered into a TLB assigned to an internal memory so that a certain memory area (page) is assigned and selected as a function of an APID.
This method yields the possibility of efficiently managing and accessing data of different applications within a VPU.
There is the option of explicitly deleting data of certain APIDs (APID-DEL) and/or writing into external (main) memories and/or other (peripheral) data streams (APID-FLUSH). This may take place whenever an application is terminated. APID-DEL and/or APID-FLUSH may be triggered by a configuration and/or by a higher-level loading unit (CT) and/or externally.
The following processing example is presented to illustrate the method.
An application Q (e.g., APID=Q) may include a configuration for reading operands (e.g., ConfigID=j), a configuration for processing operands (e.g., ConfigID=w), and a configuration for writing results (e.g., ConfigID=s).
Configuration j is executed first to read the operands chronologically optimally decoupled. Configurations of other applications may be executed simultaneously. The operands are written from external (main) memories and/or (peripheral) data streams into certain internal memories and/or memory areas according to the APID identifier.
Configuration w is executed to process the stored operands. To do so, the corresponding operands in the internal memories and/or memory areas are accessed by citation of APIDs. Results are written into internal memories and/or memory areas accordingly by citation of APIDs. Configurations of other applications may be executed simultaneously. In conclusion, configuration s writes the stored results from the internal memories and/or memory areas into external (main) memories and/or other (peripheral) data streams. Configurations of other applications may be executed simultaneously.
To this extent, the basic sequence of the method corresponds to that described above for optimization of memory access.
If data for a certain APID is not present in the memories or if there is no longer any free memory space for this data, a page fault may be triggered for transmission of the data.
While a module was initially assumed in which a field of reconfigurable elements is provided having little additional wiring, such as memories, FIFOs, and the like, it is also possible to use the ideas according to the present invention for systems known as “systems on a chip” (SoC). For SoCs the terms “internal” and “external” are not completely applicable in the traditional terminology, e.g., when a VPU is linked to other modules (e.g., peripherals, other processors, and in particular memories) on a single chip. The following definition of terms may then apply; this should not be interpreted as restricting the scope of the invention but instead is given only as an example of how the ideas of the present invention may be applied with no problem to constructs which traditionally use a different terminology:
A preferred embodiment will now be described.
In a particularly preferred embodiment, data processing PAEs are located and connected locally in the PA (e.g., ALUs, logic, etc.). RAM-PAEs may be incorporated locally into the PA, but in a particularly preferred embodiment they are remote from the PA or are placed at its edges (see, e.g., German Patent Application No. DE 100 50 442.6). This takes place so as not to interfere with the homogeneity of the PA in the case of large RAM-PAE memories, where the space required is much greater than with ALU-PAEs and because of a gate/transistor layout (e.g., GDS2) of memory cells, which usually varies greatly. If the RAM-PAEs have dedicated connections to an external bus system (e.g., global bus), they are preferably located at the edges of a PA for reasons of layout, floor plan, and manufacturing.
The configurable bus system of the PA is typically used for the physical connection.
In an expanded embodiment, PAEs and interface modules, as well as additional configurable modules, if necessary, have a dedicated connection to a dedicated global bus, e.g., a GlobalTrack.
Interface modules and in particular protocol converters are preferably remote from the PA and are placed outside of its configuration. This takes place so as not to interfere with the homogeneity of the PA and because of a gate/transistor layout (e.g., GDS2) of the interface modules/protocol converters, which usually varies greatly. In addition, the connections to external units are preferably placed at the edges of a PA for reasons of layout, floor plan, and manufacturing. The interface modules are preferably connected to the PA by the configurable bus system of the PA, the interface modules being connected to its outer edges. The bus system allows data exchange to take place configurably between interface modules and any PAEs within the PA. In other words, within one or different configurations, some interface modules may be connected to RAM-PAEs, for example, while other interface modules may be connected to ALU-PAEs, for example.
The IO-FIFOs are preferably integrated into the protocol converter. To permit a greater flexibility in the assignment of the internal data streams to the external data streams, the interface modules and protocol converters are designed separately and are connected via a configurable bus system.
The present invention is explained in greater detail below only as an example and in a nonrestrictive manner with reference to the drawings.
The individual units are interlinked via bus systems (0104). Interface modules (interface modules and protocol converters, if necessary) (0105) are located at the edges of the core and are connected to external buses (IO), as similarly described in German Patent Application No. DE 196 54 595.1. The interface modules may have different designs, depending on the implementation, and may fulfill one or more of the following functions, for example:
The diagrams in
It should be pointed out again that input of data from the input FIFO into the RAM-PAEs or 0201 and writing of data from 0201 or the RAM-PAEs may take place simultaneously.
It should likewise be pointed out that the input/output FIFOs are able to receive and/or send external data continuously during steps a-f.
Multiple configurations take place for data processing (a-e).
The data may be read into memories and/or directly (0403) into the PA from the FIFOs (input FIFOs). During the input operation, data may be written from the PA and/or memories into FIFOs (output FIFOS) (0404). For data output, data may be written from the memories and/or directly (0405) from the PA into the FIFOs. Meanwhile, new data may be written from the input FIFOs into memories and/or the PA (0406).
New data (0407) may already be entered during a last configuration, for example.
During the entire processing, data may be transmitted from externally into the input FIFOs (0408) and/or from the output FIFOs to externally (0409).
In an expanded embodiment, a PAE may additionally have a connection to a dedicated global bus (0505) (e.g., a GlobalTrack) and may thus communicate directly with a global, and if necessary also an external memory and/or peripheral unit, for example. In addition, a global bus may be designed so that different PAEs may communicate directly with one another via this bus, and in a preferred embodiment they may also communicate with modules for an external connection (e.g., interface modules). A bus system such as that described in German Patent Application No. DE 197 04 742.4, for example, may be used for such purposes.
The data processing unit (0501) may be designed for ALU-PAEs as an arithmetic logic unit (ALU), for example. Different ALU-PAEs may use different ALUs and bus connection systems. One ALU may have more than two bus connections to 0104 a and/or 0104 b, for example.
The data processing unit (0501) may be designed as a memory for RAM-PAEs, for example. Different RAM-PAEs may use different memories and bus connection systems. For example, a memory may have a plurality, in particular, more than two bus connections to 0104 a and/or 0104 b to allow access of a plurality of senders/receivers to one memory, for example. Accesses may preferably also take place simultaneously (multi-port).
The function of the memory includes, for example, the following functions or combinations thereof: random access, FIFO, stack, cache, page memory with MMU method.
In addition, in a preferred embodiment, the memory may be preloaded with data from the CT (e.g., constants, lookup tables, etc.). Likewise, in an expanded embodiment, the CT may read back data from the memory via 0504 (e.g., for debugging or for changing tasks).
In another embodiment, the RAM-PAE may have a dedicated connection (0505) to a global bus. The global bus connects a plurality of PAEs among one another and in a preferred embodiment also to modules for an external connection (e.g., interface modules). The system described in German Patent Application No. DE 197 04 742.4 may be used for such a bus system.
RAM-PAEs may be wired together in such a way that an n-fold larger memory is created from a plurality (n) of RAM-PAEs.
A bus system (0804 a, 0804 b) is provided between interface modules and/or PAEs having a dedicated global bus (0801) and protocol converters (0802). In a preferred embodiment, 0804 is able to transmit pipelined data over a plurality of register stages. 0804 a and 0804 b are interconnected via switches (e.g., 0805) which are designed as transmission gates and/or tristate buffers and/or multiplexers, for example. The multiplexers are triggered by rows and columns. Triggering units (0806) control the data transfer of the interface modules and/or PAEs having a dedicated global bus (0801) to the protocol converters (0802), i.e., in the transfer direction 0804 a to 0804 b. Triggering units (0807) control the data transfer of the protocol converters (0802) to the interface modules and/or the PAEs having a dedicated global bus (0801), i.e., in the transfer direction 0804 b to 0804 a. The triggering units (0806) each decode address areas for selection of the protocol converters (0802); the triggering units (0807) each decode IOIDs for selection of the interface modules and/or PAEs having a dedicated global bus (0801).
Triggering units may operate according to different types of triggering, e.g., fixed connection without decoding; decoding of addresses and/or IOIDs; decoding of addresses and/or IOIDs and arbitration. One or multiple data words/address words may be transmitted per arbitration. Arbitration may be performed according to different rules. The interface modules may preferably have a small FIFO for addresses and/or data in the output direction and/or input direction. A particular arbitration rule preferably arbitrates an interface module having a FULL FIFO or an EMPTY FIFO or a FIFO to be emptied (FLUSH), for example.
Triggering units may be designed as described in German Patent Application No. DE 199 26 538.0 (
The protocol converters are coupled to external bus systems (0808), a plurality of protocol converters optionally being connected to the same bus system (0809), so that they are able to utilize the same external resources.
The IO-FIFOs are preferably integrated into the protocol converters, a FIFO (BURST-FIFO) for controlling burst transfers for the external buses (0808) being connected downstream from them if necessary. In a preferred embodiment, an additional FIFO stage (SYNC-FIFO) for synchronizing the outgoing addresses with the incoming data is connected downstream from the FIFOs.
Various programmable/configurable FIFO structures are depicted in 0820-0823, where A indicates the direction of travel of an address FIFO, D indicates the direction of travel of a data FIFO. The direction of data transmission of the FIFOs depends on the direction of data transmission and the mode of operation. If a VPU is operating as a bus master, then data and addresses are transmitted from internally to the external bus in the event of a write access (0820), and in the event of a read access (0821) addresses are transmitted from internally to externally and data from externally to internally.
If a VPU is operating as a bus slave, then data and addresses are transmitted from the external bus to internally in the event of a write access (0822) and in the event of a read access (0823) addresses are transmitted from externally to internally and data is transmitted from internally to externally.
In all data transfers, addresses and/or data and/or IOIDs and/or APIDs may be assigned and also stored in the FIFO stages.
In a particularly preferred embodiment, the transfer rate (operating frequency) of the bus systems 0104, 0804, and 0808/0809 may each be different due to the decoupling of the data transfers by the particular FIFO stages. In particular the external bus systems (0808/0809) may operate at a higher transfer rate, for example, than the internal bus systems (0104) and/or (0804).
Addresses (preferably identifiers, e.g., with IOIDs and/or APIDs) are transmitted via internal bus system 0104 to interface modules and/or PAEs having a dedicated global bus, which preferably have an internal FIFO (0901). The addresses are transmitted to an IO-FIFO (0903) via a bus system (e.g., 0804) which preferably operates as a pipeline (0902). The addresses are transmitted to a BURST-FIFO (0905) via another bus (0904) which may be designed as a pipeline but which is preferably short and local. The BURST-FIFO ensures correct handling of burst transfers via the external bus system, e.g., for controlling burst addresses and burst sequences and repeating burst cycles when errors occur. IOIDs and/or APIDs of addresses (0906) which are transmitted via the external bus system may be transmitted together with the addresses and/or stored in an additional SYNC-FIFO (0907). The SYNC-FIFO compensates for the latency between the outgoing address (0906) and the incoming data (0909). Incoming data may be assigned IOIDs and/or APIDs (0908) of the addresses referencing them via the SYNC-FIFO (0910). Data (and preferably IOIDs and/or APIDs) is buffered in an IO-FIFO (0911) and is subsequently transmitted via a bus system (e.g., 0804), which preferably functions as a pipeline (0912), to an interface module and/or PAE having a dedicated global bus (0913), preferably including an internal FIFO. Data is transmitted from here to the internal bus system (0104).
Instead of to the IO-FIFO (0911), incoming data may optionally be directed first to a second BURST-FIFO (not shown), which behaves like BURST-FIFO 0905 if burst-error recovery is also necessary in read accesses. Data is subsequently relayed to 0911.
Multiple collectors may be implemented for which multiple triggering units 1006 and 1007 are used.
A collector may be segmented into multiple memory areas. Each memory area may operate independently in different memory modes, e.g., as random access memory, FIFO, cache, MMU page, etc.
A translation table (TLB) (1002) may be assigned to a collector to permit an MMU-type mode of operation. Page management may function, e.g., on the basis of segment addresses and/or other identifiers, e.g., APIDs and/or IOIDs.
A DMA or multiple DMAs are preferably assigned to a collector to perform data transfers with external (main) memories and/or other (peripheral) data streams, in particular to automatically permit the MMU function of page management (loading, writing). DMAs may also access the TLB for address translation between external (main) memories and/or other (peripheral) data streams and collector. In one possible mode of operation, DMAs may receive address specifications from the array (PA), e.g., via 0804.
DMAs may be triggered by one or more of the following units: an MMU assigned to the collector, e.g., in the case of page faults; the array (PA); an external bus (e.g., 0809); an external processor; a higher-level loading unit (CT).
Collectors may have access to a dedicated bus interface (1004), preferably DMA-controlled and preferably master/slave capable, including a protocol converter, corresponding to or similar to protocol converters 0802 having access to external (main) memories and/or other (peripheral) data streams.
An external processor may have direct access to collectors (1007).
The channels preferably function in a FIFO-like mode of operation to perform data transfers with 1208.
In the example depicted here, a channel (1207), which has been filled by a previous configuration or application, is still being written to 1208 during data processing within 1201 described here. This channel may also include internal FIFO stages and/or PAE-RAMs/PAE-RAM pages and/or collectors/collector pages. The addresses may be computed currently by a configuration (OADR-CONF) running in parallel in 1201 and/or computed in advance and/or computed by DMA operations of a (1003). In particular, an address computation within 1201 (OADR-CONF) may be sent to a collector or its DMA to address and control the data transfers of the collector.
In addition, data for a subsequent configuration or application is simultaneously loaded into another channel (1206). This channel too may include internal FIFO stages and/or PAE-RAMs/PAE-RAM pages and/or collectors/collector pages. The addresses may be computed currently by a configuration (IADR-CONF) running in parallel in 1201 and/or computed in advance and/or computed by DMA operations of a (1003). In particular, an address computation within 1201 (IADR-CONF) may be sent to a collector or its DMA to address and control the data transfers of the collector. Individual entries into the particular channels may have different identifiers, e.g., IOIDs and/or APIDs, enabling them to be assigned to a certain resource and/or memory location.
The function of an output FIFO which transmits its values to a burst-capable bus (BBUS) is to be described first. A first pointer (1301) points to the data entry within a memory (1304) currently to be output to the BBUS. With each data word output (1302), 1301 is moved by one position. The value of pointer 1301 prior to the start of the current burst transfer has been stored in a register (1303). If an error occurs during the burst transfer, 1301 is reloaded with the original value from 1303 and the burst transfer is restarted.
A second pointer (1305) points to the current data input position in the memory (1304) for data to be input (1306). To prevent overwriting of any data still needed in the event of an error, pointer 1305 is compared (1307) with register 1303 to indicate that the BURST-FIFO is full. The empty state of the BURST-FIFO may be ascertained by comparison (1308) of the output pointer (1301) with the input pointer (1305).
If the BURST-FIFO operates for input data from a burst transfer, the functions change as follows:
1301 becomes the input pointer for data 1306. If faulty data has been transmitted during the burst transfer, the position prior to the burst transfer is stored in 1303. If an error occurs during the burst transfer, 1301 is reloaded with the original value from 1303 and the burst transfer is restarted.
The pointer points to the readout position of the BURST-FIFO for reading out the data (1302). To prevent premature readout of data of a burst transfer that has not been concluded correctly, 1305 is compared with the position stored in 1303 (1307) to indicate an empty BURST-FIFO. A full BURST-FIFO is recognized by comparison (1308) of input pointer 1301 with the output pointer (1305).
The address value (1313) of a first access is stored in a register (1310). The address value of a subsequent data transfer is compared (1312) with the address value (1311) of 1310, which has been incremented by the address difference between the first data transfer and the second data transfer of the burst transfer (typically one word wide). If the two values are the same, then the difference between the first address and the second address corresponds to the address difference of the burst transfer between two burst addresses. Thus, this is a correct burst. If the values are not the same, the burst transfer must be aborted.
The last address (1313) checked (the second address in the writing) is stored in 1310 and then compared with the next address (1313) accordingly.
To ascertain whether the burst limits (boundaries) have been maintained, the address bit(s) at which the boundary of the current address value (1313) is located is (are) compared with the address bits of the preceding address value (1310) (e.g., XOR 1314). If the address bits are not the same, the boundary has been exceeded and the control of the burst must respond accordingly (e.g., termination of the burst transfer and restart).
Write data (1401) is preferably sent to the memories via pipeline stages (1402). Read data (1403) is preferably removed from the memories also via pipeline stages (1404). Pipeline stage 1404 includes a multiplexer, which forwards the particular active data path. The active data path may be recognized, for example, by a RDY handshake applied.
A unit (RangeCheck, 1405) for monitoring the addresses (1406) for correct values within the address space may optionally be provided.
For example, FREGs of the PAEs according to
Any other embodiments and combinations of the present inventions described here are possible and are self-evident in view of the foregoing, to those skilled in the art.